andrew: charriso

Computational Photography - Fall 2011

Assignment 5 (PART 1 & 2)



Part 1

I found it very difficult to get good alignment points. Even single pixel inaccuracies seem to have a big effect. Consequently, some of the seems are visible.

Implemented linear blending (first three panoramas use this).

Use a click based interface for selecting alignment points (see first image with black background).

See bells and whistles section for examples of simple image warps.

For bells and whistles:
Plastered Efros' face onto road signs and house windows
Put "EFROS" onto street names.
Graffiti (masked) onto buildings.


Source:

Simple first test (selection interface shown, as well as the linear blend mask applied to the image on the right):


Real Panorama 1:

Source:

Result:

 


Real Panorama 2:

Source:

Result:


Bells and Whistles

= +


=

+


Source:

Result: (there are 5 graffiti marks in the scene)



Part 2

Linear blending on left/right edges.
Implemented Lowe's "should I marry" algorithm.
Points must agree they are each others best match (mutual agreement)
Implemented non-maximal suppression for better spatial distribution of interest points.
Use a 4-point RANSAC (10,000 cycles)



Example 8x8 patches (subsampled from 40x40 patches)



Initial tests:

  


Final output (3 mosaics):
Correspondance points shown in red, yellow lines connecting mutual best matches.
Yellow and cyan circles show best homography points for center and right images.

Automatic:
Manual:


Automatic:
Manual:


Automatic: Manual:




Part 3

Tour into the Picture:  Single View Modeling

High res images can be found here




Orig (left), interface with vanishing point lines (right)

     




Original Image

 
 




(Buckingham Palace)